Anders Sandberg: revision 1 to 2

Revision 1

Revision 2

categories: ["people", "minds"]

tags: ["whole brain emulation", "existential risk", "forecasting", "computational neuroscience", "transhumanism"]

summary: "Swedish computational neuroscientist and futures researcher who co-wrote the whole brain emulation roadmap and works on the physical limits of long-term civilization."

summary: "Swedish computational neuroscientist and futures researcher who co-wrote the whole brain emulation roadmap and works on the physical limits of long-term civilization."

updated: "2026-07-27"

updated: "2026-10-29"

issues: ["The anthropic shadow and Converging Cognitive Enhancements papers are discussed without citations."]

---

## Career

Sandberg's doctorate, completed at Stockholm University in 2002, was on attractor neural network

models of memory — the mathematics of how a network of units settles into stable states that behave

models of memory, the mathematics of how a network of units settles into stable states that behave

like stored patterns. That background explains his position on emulation: he approaches the question

like stored patterns. That background explains his position on emulation. He approaches the question

as a modeller who has built neural systems, not as a philosopher arguing from the outside.

as a modeller who has built neural systems rather than as a philosopher arguing from the outside.

He chaired the Swedish transhumanist association during the 1990s and co-founded a Stockholm think

tank before joining the Future of Humanity Institute in 2006, where he worked with

tank before joining the Future of Humanity Institute in 2006. There he worked with [[nick-bostrom]],

[[nick-bostrom]], Toby Ord, [[eric-drexler]] and others until the institute closed in 2024. He

Toby Ord, [[eric-drexler]] and others until the institute closed in 2024, after which he moved to

subsequently moved to the Institute for Futures Studies in Stockholm. He is a prolific public

the Institute for Futures Studies in Stockholm. He is a prolific public communicator, and much of

communicator, and much of his output takes the form of estimates published as blog posts, technical

his output appears as estimates in blog posts, technical reports and talks rather than as

reports and talks rather than as conventional papers.

conventional papers.

## Whole brain emulation

The 2008 report written with Bostrom is the reference document for the emulation

question.[^roadmap2008] Its method is to refuse to argue about whether emulation is possible and

question.[^roadmap2008] It declines to argue about whether emulation is possible. Instead it asks

instead ask what would have to be true at each level of description, then estimate the requirements

what would have to be true at each level of description, then estimates the requirements under each

under each assumption.

assumption.

Its premise, stated as a premise rather than defended, is [[substrate-independence]]: that what

matters about a brain is the organization of its causal structure rather than the material

implementing it.

The roadmap defines a ladder of models: from a coarse brain-region simulation, through spiking

The roadmap defines a ladder of models, from a coarse brain-region simulation through spiking neural

neural networks, compartment models with realistic dendrites, molecular-level channel dynamics, and

networks, compartment models with realistic dendrites and molecular-level channel dynamics, down to

down to whole-molecule and quantum descriptions. Each rung carries an estimate of storage, compute

whole-molecule and quantum descriptions. Each rung carries an estimate of storage, compute and

and scanning resolution. The estimates span many orders of magnitude, and the report is explicit

scanning resolution. The estimates span many orders of magnitude, and the report says explicitly

that the correct rung is unknown — this is the central finding, not a hedge. It also identifies

that the correct rung is unknown; that uncertainty is its central finding, not a hedge. It also

scanning throughput rather than compute as the likely binding constraint, a judgement that

identifies scanning throughput rather than compute as the likely binding constraint, a judgement

[[connectomics]] has broadly borne out: reconstructing a cubic millimetre of cortex remains a

that [[connectomics]] has broadly borne out, since reconstructing a cubic millimetre of cortex

multi-year effort for a large collaboration. The same analysis makes structural

remains a multi-year effort for a large collaboration. The same analysis makes structural

[[brain-preservation]] a coherent research target, since a preserved connectome is a scannable

[[brain-preservation]] a coherent research target, because a preserved connectome is a scannable

object even if nobody can yet scan it at scale.

> [!key] What the roadmap actually establishes

> Not that emulation will work. It establishes that the requirements are finite and estimable *given

> The roadmap does not show that emulation will work. It establishes that the requirements are

> a level of description*, and that the disagreement between optimists and sceptics is almost

> finite and estimable *given a level of description*, and that the disagreement between optimists

> entirely about which level suffices. That reframing is the report's contribution.

> and sceptics is almost entirely about which level suffices. That reframing is the report's

> contribution.

Sandberg has separately written on the ethics of emulation — the welfare of a running emulation, the

Sandberg has separately written on the ethics of emulation: the welfare of a running emulation, the

implications of copying, and the research ethics of experiments on emulated brains, which is

implications of copying, and the research ethics of experiments on emulated brains, which amounts to

[[machine-consciousness]] with the moral stakes made unavoidable. He is consistently more cautious

[[machine-consciousness]] with the moral stakes made unavoidable. In public he is consistently more

in public than the [[mind-uploading]] discourse that cites him.

cautious than the [[mind-uploading]] discourse that cites him.

## Quantifying the long term

A recurring pattern in Sandberg's work is taking a speculative claim and computing its physical

Sandberg's work repeatedly takes a speculative claim and computes its physical requirements. With

requirements. With Stuart Armstrong he calculated the energy and engineering cost of launching

Stuart Armstrong he calculated the energy and engineering cost of launching self-replicating probes

self-replicating probes to every reachable galaxy, concluding that intergalactic settlement is

to every reachable galaxy. They concluded that intergalactic settlement is within the physical means

within the physical means of a civilization not much more capable than a plausible near-future

of a civilization not much more capable than a plausible near-future one, a result that sharpens the

one — which sharpens rather than resolves the Fermi paradox.[^eternity2013] With Drexler and Ord he

Fermi paradox without resolving it.[^eternity2013] With Drexler and Ord he argued that the paradox

argued that the paradox partly dissolves under correct handling of uncertainty: multiplying point

partly dissolves when uncertainty is handled correctly. Multiplying point estimates in the Drake

estimates in the Drake equation discards the distribution, and propagating the actual uncertainty

equation discards the distribution, and propagating the actual uncertainty places substantial

places substantial probability mass on humanity being alone.[^dissolving2018]

probability mass on humanity being alone.[^dissolving2018]

He has written on the thermodynamic limits of computation, on "Jupiter brains" as the physical

ceiling for a single computational structure, and on the aestivation hypothesis — the proposal that

ceiling for a single computational structure, and on the aestivation hypothesis, the proposal that

an advanced civilization might wait for the cosmic background to cool before computing, because

computation is cheaper at low temperature. With Ord and others he formulated the argument that

computation is cheaper at low temperature. With Ord and others he formulated the argument that model

model uncertainty dominates very low probability estimates: a calculation giving a one-in-a-billion

uncertainty dominates very low probability estimates: a calculation giving a one-in-a-billion chance

chance of catastrophe is worthless if the chance the calculation is wrong is one in a

of catastrophe is worthless if the chance that the calculation is wrong is one in a

thousand.[^probing2010]

## Enhancement and risk

Sandberg's enhancement work is mostly with Bostrom. "Converging Cognitive Enhancements" surveys the

routes — pharmacological, genetic, external, and interface-based — and argues that they interact

pharmacological, genetic, external and interface-based routes and argues that they interact rather

rather than compete. The "wisdom of nature" heuristic asks why evolution has not already made a

than compete. The "wisdom of nature" heuristic asks why evolution has not already made a proposed

proposed improvement, and treats a satisfactory answer as a precondition for expecting the

improvement, and treats a satisfactory answer as a precondition for expecting the improvement to be

improvement to be free of hidden costs; the usual acceptable answers are changed environments,

free of hidden costs. The usual acceptable answers are changed environments, evolutionary

evolutionary constraints, and value discordance between fitness and welfare. It is a rare piece of

constraints, and value discordance between fitness and welfare. Pro-enhancement writing rarely

pro-enhancement writing that supplies a filter its own conclusions must pass. See

supplies a filter that its own conclusions must pass, and this heuristic does. See

[[human-enhancement]] and [[nootropics]].

On risk, the "anthropic shadow" argument he co-wrote holds that the historical record systematically

understates the frequency of catastrophes severe enough to have prevented observers from existing —

understates the frequency of catastrophes severe enough to have prevented observers from existing.

so the empirical base rate for some [[existential-risk|extinction-level events]] is biased downward

The empirical base rate for some [[existential-risk|extinction-level events]] is therefore biased

by the fact that anyone is around to compute it.

downward by the fact that anyone is around to compute it.

## Reception and legacy

Sandberg is cited across futures studies, AI safety and philosophy, and is unusual in retaining

Sandberg is cited across futures studies, AI safety and philosophy, and he is unusual in keeping his

credibility with mainstream scientists while working on subjects they generally avoid. The reason is

methodological: he states assumptions, publishes the arithmetic, and marks the gap between what

his method: he states assumptions, publishes the arithmetic, and marks the gap between what follows

follows from physics and what follows from speculation. Criticism concentrates on the limits of that

from physics and what follows from speculation. Criticism concentrates on the limits of that method.

method — that computing the physical requirements of a scenario says nothing about its biological or

Computing the physical requirements of a scenario says nothing about its biological or social

social plausibility, and that the roadmap's neutral framing gives emulation more credibility than

plausibility, critics argue, and the roadmap's neutral framing gives emulation more credibility than

neuroscience currently warrants.

His long-announced book on grand futures, covering the physical possibilities open to a

His long-announced book on grand futures, covering the physical possibilities open to a civilization

civilization over cosmological time, has circulated for years mainly as drafts and lecture

over cosmological time, has circulated for years mainly as drafts and lecture material, and is

material, and is easier to encounter in talks than in print. Its subject is the same one this

easier to encounter in talks than in print. Its subject is the one this wiki's

wiki's [[future-of-humanity]] article takes up, approached from the side of physical limits rather

[[future-of-humanity]] article takes up, approached from physical limits rather than from

than of technology. And its central

technology. Its central question is one his quantitative work keeps raising and cannot answer. The

question is the one his quantitative work keeps producing and cannot answer: the physics permits

physics permits enormously more than humanity currently does, so the binding constraint is not

enormously more than humanity currently does, so the binding constraint is not physical, and

physical, and nothing in the calculations says what it is instead.

nothing in the calculations says what it is instead.

## See also